Tencent Hunyuan Open-Sources WorldCompass, a Reinforcement Learning Framework for World Models, Boosting Complex Action Interaction Accuracy by Over 35%

Tencent Hunyuan 3D team announced the open-source of WorldCompass, the industry's first reinforcement learning post-training framework for world models, which is the official RL extension module of Hunyuan World Model 1.5.

WorldCompass is designed for long-sequence, interactive world models, acting like a 'compass' to guide models to follow user instructions more accurately and maintain long-term visual consistency.

Experimental verification shows that this framework significantly improves the interaction accuracy and visual fidelity of the open-source model WorldPlay: in complex combined action scenarios, accuracy increases from ~20% to ~55% (over 35% improvement), basic action accuracy increases by ~10%; it also achieves better results in Stanford's WorldScore benchmark. Relevant code and model details have been open-sourced.